lkml.org 
[lkml]   [2018]   [Aug]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [RFC PATCH v3 12/24] x86/mm: Modify ptep_set_wrprotect and pmdp_set_wrprotect for _PAGE_DIRTY_SW
On Thu, Aug 30, 2018 at 6:09 PM Dave Hansen <dave.hansen@linux.intel.com> wrote:
>
> On 08/30/2018 08:49 AM, Jann Horn wrote:
> >> @@ -1203,7 +1203,28 @@ static inline pte_t ptep_get_and_clear_full(struct mm_struct *mm,
> >> static inline void ptep_set_wrprotect(struct mm_struct *mm,
> >> unsigned long addr, pte_t *ptep)
> >> {
> >> + pte_t pte;
> >> +
> >> clear_bit(_PAGE_BIT_RW, (unsigned long *)&ptep->pte);
> >> + pte = *ptep;
> >> +
> >> + /*
> >> + * Some processors can start a write, but ending up seeing
> >> + * a read-only PTE by the time they get to the Dirty bit.
> >> + * In this case, they will set the Dirty bit, leaving a
> >> + * read-only, Dirty PTE which looks like a Shadow Stack PTE.
> >> + *
> >> + * However, this behavior has been improved and will not occur
> >> + * on processors supporting Shadow Stacks. Without this
> >> + * guarantee, a transition to a non-present PTE and flush the
> >> + * TLB would be needed.
> >> + *
> >> + * When change a writable PTE to read-only and if the PTE has
> >> + * _PAGE_DIRTY_HW set, we move that bit to _PAGE_DIRTY_SW so
> >> + * that the PTE is not a valid Shadow Stack PTE.
> >> + */
> >> + pte = pte_move_flags(pte, _PAGE_DIRTY_HW, _PAGE_DIRTY_SW);
> >> + set_pte_at(mm, addr, ptep, pte);
> >> }
> > I don't understand why it's okay that you first atomically clear the
> > RW bit, then atomically switch from DIRTY_HW to DIRTY_SW. Doesn't that
> > mean that between the two atomic writes, another core can incorrectly
> > see a shadow stack?
>
> Good point.
>
> This could result in a spurious shadow-stack fault, or allow a
> shadow-stack write to the page in the transient state.
>
> But, the shadow-stack permissions are more restrictive than what could
> be in the TLB at this point, so I don't think there's a real security
> implication here.

How about this:

Three threads (A, B, C) run with the same CR3.

1. a dirty+writable PTE is placed directly in front of B's shadow stack.
(this can happen, right? or is there a guard page?)
2. C's TLB caches the dirty+writable PTE.
3. A performs some syscall that triggers ptep_set_wrprotect().
4. A's syscall calls clear_bit().
5. B's TLB caches the transient shadow stack.
[now C has write access to B's transiently-extended shadow stack]
6. B recurses into the transiently-extended shadow stack
7. C overwrites the transiently-extended shadow stack area.
8. B returns through the transiently-extended shadow stack, giving
the attacker instruction pointer control in B.
9. A's syscall broadcasts a TLB flush.

Sure, it's not exactly an easy race and probably requires at least
some black timing magic to exploit, if it's exploitable at all - but
still. This seems suboptimal.

> The only trouble is handling the spurious shadow-stack fault. The
> alternative is to go !Present for a bit, which we would probably just
> handle fine in the existing page fault code.

\
 
 \ /
  Last update: 2018-08-30 20:09    [W:0.067 / U:4.968 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site